Facial symmetry has the appeal of a clean rule. Match the left and right halves more closely, the story goes, and a face should be judged more attractive. The idea has appeared in evolutionary accounts, beauty advice and experiments built from digitally altered portraits.
A large cross-cultural analysis complicates it. Karel Kleisner and colleagues measured 1,550 natural faces from ten countries and found no robust independent relationship between symmetry and attractiveness once other properties of facial shape were considered. What did predict higher ratings, consistently across the sampled populations, was being closer to the average face for the same sex and population.
This is not evidence that everyone wants the same face. Nor does it show that symmetry can never influence perception. The result is more precise: in these photographs, asymmetry did not explain attractiveness ratings after the model separated it from distinctiveness and sex-typicality.
Ten countries, 1,550 faces and a controlled comparison
The study in Evolution and Human Behavior assembled standardized frontal photographs from Brazil, Cameroon, Colombia, Czechia, India, Namibia, Romania, Turkey, the United Kingdom and Vietnam. In all, the dataset contained 1,550 faces.
Most of the photographs had been gathered for earlier research, with Indian and Vietnamese samples added for this analysis. The pictures were made with digital cameras, external lighting and plain white or grey backgrounds. Standardization reduces distractions such as angle and scenery, although it cannot turn photographs collected in different projects into one perfectly uniform dataset.
Attractiveness was defined as an average rating on a seven-point scale. The main analysis used judgments from opposite-sex raters drawn from the same culture as the photographed person. That makes the outcome a local social perception, not a universal property measured by an instrument.
The team placed 72 landmarks and semilandmarks on each face. These points traced stable locations and curves around the eyes, nose, mouth, jaw and other parts of the facial outline. Geometric morphometrics then allowed the researchers to compare configurations while accounting for position, rotation and overall scale.
What “average” means in this study
In everyday speech, average can sound like a verdict: ordinary, unexceptional or forgettable. Here it was a geometric quantity. The researchers calculated a mean facial shape separately for each sex within each sampled country, then measured how far each individual face lay from that mean.
A shorter distance meant the landmark configuration was more average. A longer distance meant it was more distinctive. A Brazilian man was therefore compared with the relevant Brazilian male mean, and a Vietnamese woman with the relevant Vietnamese female mean. The analysis did not construct a single global template and declare it ideal.
Nor did it measure every feature that might make a face seem familiar. Skin colour and texture, hairstyle, expression, age cues and grooming are not contained fully in a map of facial landmarks. “Closer to average” in the result means closer in the shape space the researchers defined.
This distinction matters because averageness has several possible explanations. Familiar or prototype-like shapes may be easier for visual systems to process. Average configurations may also contain fewer unusual deviations. The analysis shows an association with ratings; it cannot decide which mechanism caused it.
How symmetry becomes tangled with distinctiveness
Symmetry was measured separately. The researchers reflected each facial configuration and calculated the distance between the original and mirrored shapes. Greater distance represented greater asymmetry.
The complication is that an asymmetric feature normally changes two variables at once. It makes the two sides less alike, but it also pushes the whole configuration farther from the group mean. In this dataset, asymmetry and distinctiveness correlated at 0.29 for women’s faces and 0.27 for men’s faces.
Those are not overwhelming relationships, but they are large enough to confuse a simple comparison. If less symmetrical faces receive lower ratings, symmetry may be responsible. Or their greater distinctiveness may be responsible. Looking at one predictor without the other cannot cleanly distinguish the two.
This is also a problem for digitally manipulated photographs. Making a face more symmetrical can simultaneously make it more typical, smooth out natural irregularities or create subtle changes beyond the intended manipulation. A 2021 review found that preferences for facial symmetry depend partly on study design, including whether researchers compare real faces, mirrored composites or modified images.
The model separated the travelling companions
Kleisner’s team used Bayesian multilevel models to estimate the relationships of distinctiveness, asymmetry and sex-typicality with attractiveness at the same time. The models allowed slopes and starting points to vary by country rather than assuming every sampled population produced an identical pattern.
Once these features entered together, distinctiveness had a clear negative relationship with attractiveness for female and male faces. Faces farther from their sex- and country-specific mean tended to receive lower ratings. The direction was consistent across all ten sampled populations.
Asymmetry did not show the expected independent penalty. Its estimated slopes sat close to zero for men’s faces and were slightly positive for women’s faces after distinctiveness was held constant. That does not mean asymmetry made women more attractive; the key point is that the analysis found no robust evidence for the predicted negative effect.
The researchers also measured sex-typicality by locating facial shapes along the difference between the average male and female configurations. Greater femininity predicted higher ratings for women’s faces. Greater masculinity did not provide a clear equivalent advantage for men’s faces.
Adding body mass index and height for the subset of 1,106 people with those data did not overturn the central pattern. Analyses that accounted for age, and models run at the level of individual ratings rather than face averages, likewise served as robustness checks rather than producing a different story.
No independent effect is not the same as no effect ever
The wording matters. An “independent effect” here means an association left over after the statistical model holds the other measured facial characteristics constant. It does not mean symmetry has been proved irrelevant to every face, every viewer and every circumstance.
Very pronounced asymmetry might matter in ways not captured well by a mostly ordinary sample. Symmetry might affect judgments under another viewing angle, in motion or alongside texture and colour. A landmark analysis can also miss smaller surface features that observers notice.
The study is observational. It describes how naturally varying faces and ratings were related; it did not randomly assign people to have different facial structures. Statistical adjustment helps disentangle correlated traits, but it is not the same as a controlled causal intervention, and no model contains every characteristic an observer may use.
The result therefore narrows the symmetry claim rather than abolishing it. Among these natural faces, the evidence did not support symmetry as a separate general ingredient of attractiveness. That is a substantial correction to the simple rule, but not a declaration that bilateral balance is perceptually invisible.
A broad sample is not all of humanity
Ten countries offer a far wider test than the single-university samples common in face-perception research. They also span different regions and populations. Still, “across ten countries” should not be read as ten nationally representative surveys.
The faces and raters came from particular research samples. Country labels can contain enormous internal diversity, and a culture is not reducible to a national border. Opposite-sex ratings also leave out same-sex attraction and other forms of attraction that do not fit the study’s design.
Standardized still portraits omit much of what shapes an impression in life: movement, voice, expression, clothing, familiarity, humour and behaviour. ScienceBlog has previously covered how face-based impressions can form within one-tenth of a second. Their speed does not make them a complete or accurate reading of a person.
The same caution applies here. An attractiveness rating is a response made by an observer under specified conditions. It is not a diagnosis of health, genetic quality, character or human worth. The study measured a pattern in perception, not a hierarchy of people.
A later replication found the same basic pattern
A 2025 replication in Scientific Reports returned to the distinction between symmetry and averageness. It used 100 white faces, evenly divided between women and men, and ratings from 400 heterosexual participants.
Greater distinctiveness again predicted lower attractiveness ratings. Asymmetry did not have a significant independent relationship once the variables were considered together. Femininity predicted ratings of women’s faces, while masculinity did not clearly predict ratings of men’s faces.
The replication is useful because a second dataset reproduced the main statistical pattern. It is narrower than the original ten-population analysis, however, and cannot by itself settle how the relationships operate across ages, identities, viewing conditions and cultures.
The fairest conclusion is not that symmetry was an illusion or that averageness is a new universal beauty formula. It is that the traits were partly entangled, and a more careful model changed which one appeared to carry the signal.
Symmetry remains an intuitively attractive explanation because the eye can spot it and language can describe it in one word. Facial averageness is harder to notice directly because it is defined relative to many other faces. In this research, the less visible statistical relationship turned out to be the more consistent one.